Skip to contentSkip to conversation

KCSAN: data-race in exit_signals / prepare_signal

PATCH In review syzbot 5 participants 18 messages Oct 21, 2019 – Oct 28, 2019

Subsystems (2)
lore
Share & tools
lore

syzbot

author syzbot
#0 lore raw

Hello,

syzbot found the following crash on:

HEAD commit:    d724f94f x86, kcsan: Enable KCSAN for x86
git tree:       https://github.com/google/ktsan.git kcsan
console output: https://syzkaller.appspot.com/x/log.txt?x=13eab79f600000
kernel config:  https://syzkaller.appspot.com/x/.config?x=c0906aa620713d80
dashboard link: https://syzkaller.appspot.com/bug?extid=492a4acccd8fc75ddfd0
compiler:       gcc (GCC) 9.0.0 20181231 (experimental)

Unfortunately, I don't have any reproducer for this crash yet.

IMPORTANT: if you fix the bug, please add the following tag to the commit:

Reported-by [email protected]

==================================================================

BUG: KCSAN: data-race in exit_signals / prepare_signal

read to 0xffff888103566064 of 4 bytes by interrupt on cpu 0:

 sig_task_ignored kernel/signal.c:94 [inline]
 sig_ignored kernel/signal.c:119 [inline]
 prepare_signal+0x1f5/0x790 kernel/signal.c:956
 send_sigqueue+0xc1/0x4b0 kernel/signal.c:1859
 posix_timer_event kernel/time/posix-timers.c:328 [inline]
 posix_timer_fn+0x10d/0x230 kernel/time/posix-timers.c:354
 __run_hrtimer kernel/time/hrtimer.c:1389 [inline]
 __hrtimer_run_queues+0x288/0x600 kernel/time/hrtimer.c:1451
 hrtimer_interrupt+0x22a/0x480 kernel/time/hrtimer.c:1509
 local_apic_timer_interrupt arch/x86/kernel/apic/apic.c:1110 [inline]
 smp_apic_timer_interrupt+0xdc/0x280 arch/x86/kernel/apic/apic.c:1135
 apic_timer_interrupt+0xf/0x20 arch/x86/entry/entry_64.S:830
 arch_local_irq_enable arch/x86/include/asm/paravirt.h:778 [inline]
 __raw_spin_unlock_irq include/linux/spinlock_api_smp.h:168 [inline]
 _raw_spin_unlock_irq+0x4e/0x80 kernel/locking/spinlock.c:199
 spin_unlock_irq include/linux/spinlock.h:388 [inline]
 get_signal+0x1f4/0x1320 kernel/signal.c:2707
 do_signal+0x3b/0xc00 arch/x86/kernel/signal.c:815
 exit_to_usermode_loop+0x250/0x2c0 arch/x86/entry/common.c:159
 prepare_exit_to_usermode arch/x86/entry/common.c:194 [inline]
 syscall_return_slowpath arch/x86/entry/common.c:274 [inline]
 do_syscall_64+0x2d7/0x2f0 arch/x86/entry/common.c:299
 entry_SYSCALL_64_after_hwframe+0x44/0xa9

write to 0xffff888103566064 of 4 bytes by task 7604 on cpu 1:

 exit_signals+0x13b/0x490 kernel/signal.c:2822
 do_exit+0x1af/0x18e0 kernel/exit.c:825
 do_group_exit+0xb4/0x1c0 kernel/exit.c:983
 __do_sys_exit_group kernel/exit.c:994 [inline]
 __se_sys_exit_group kernel/exit.c:992 [inline]
 __x64_sys_exit_group+0x2e/0x30 kernel/exit.c:992
 do_syscall_64+0xcf/0x2f0 arch/x86/entry/common.c:296
 entry_SYSCALL_64_after_hwframe+0x44/0xa9

Reported by Kernel Concurrency Sanitizer on:

CPU: 1 PID: 7604 Comm: syz-executor.4 Not tainted 5.3.0+ #0
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011
==================================================================

---
This bug is generated by a bot. It may contain errors. See https://goo.gl/tpsmEJ for more information about syzbot. syzbot engineers can be reached at [email protected].

syzbot will keep track of this bug report. See:
https://goo.gl/tpsmEJ#status for how to communicate with syzbot.

#1 lore raw

[+Cc Will]

Quoted 11 lines · syzbot

Hello,

syzbot found the following crash on:

HEAD commit:    d724f94f x86, kcsan: Enable KCSAN for x86
git tree:       https://github.com/google/ktsan.git kcsan
console output: https://syzkaller.appspot.com/x/log.txt?x=13eab79f600000
kernel config:  https://syzkaller.appspot.com/x/.config?x=c0906aa620713d80
dashboard link: https://syzkaller.appspot.com/bug?extid=492a4acccd8fc75ddfd0
compiler:       gcc (GCC) 9.0.0 20181231 (experimental)
Quoted 7 lines

Unfortunately, I don't have any reproducer for this crash yet.

IMPORTANT: if you fix the bug, please add the following tag to the commit:

Reported-by [email protected]

==================================================================

BUG: KCSAN: data-race in exit_signals / prepare_signal

This traces back to Oleg fixing a race between a group stop and a thread exiting before it notices that it has a pending signal or is in the middle of do_exit() already, causing group stop to get wacky.
The original commit to fix this race is
commit d12619b5ff56 ("fix group stop with exit race") which took sighand lock before setting PF_EXITING on the thread.

Later on in
commit 5dee1707dfbf ("move the related code from exit_notify() to exit_signals()") an improvement was made for the single-threaded case and the
case where the group stop is already in progress. This removed the sighand lock around the PF_EXITING assignment.

If the race really matters and given how tsk->flags is currently accessed everywhere the simple fix for now might be:

kernel/signal.c +2 -0
index c4da1ef56fdf..cf61e044c4cc 100644
@@ -2819,7 +2819,9 @@ void exit_signals(struct task_struct *tsk)
        cgroup_threadgroup_change_begin(tsk);
 
        if (thread_group_empty(tsk) || signal_group_exit(tsk->signal)) {
+               spin_lock_irq(&tsk->sighand->siglock);
                tsk->flags |= PF_EXITING;
+               spin_unlock_irq(&tsk->sighand->siglock);
                cgroup_threadgroup_change_end(tsk);
                return;
        }

Christian

#2 lore raw
Quoted 7 lines · Christian Brauner

This traces back to Oleg fixing a race between a group stop and a thread exiting before it notices that it has a pending signal or is in the middle of do_exit() already, causing group stop to get wacky.
The original commit to fix this race is
commit d12619b5ff56 ("fix group stop with exit race") which took sighand lock before setting PF_EXITING on the thread.

Not really... sig_task_ignored() didn't check task->flags until the recent 33da8e7c81 ("signal: Allow cifs and drbd to receive their terminating signals"). But I think this doesn't matter, see below.

Quoted 8 lines

If the race really matters and given how tsk->flags is currently accessed everywhere the simple fix for now might be:

kernel/signal.c +0 -0
index c4da1ef56fdf..cf61e044c4cc 100644

kernel/signal.c void exit_signals(struct task_struct *tsk) · line 2824

Quoted 6 lines
kernel/signal.c +2 -0
@@ -2819,7 +2819,9 @@ void exit_signals(struct task_struct *tsk)
        cgroup_threadgroup_change_begin(tsk);
 
        if (thread_group_empty(tsk) || signal_group_exit(tsk->signal)) {
+               spin_lock_irq(&tsk->sighand->siglock);
                tsk->flags |= PF_EXITING;
+               spin_unlock_irq(&tsk->sighand->siglock);

Well, exit_signals() tries to avoid ->siglock in this case....

But this doesn't matter. IIUC the problem is not that exit_signals() sets PF_EXITING, the problem is that it writes to tsk->flags and kasan detects the data race.

For example, freezable_schedule() which sets PF_FREEZER_SKIP can equally "race" with sig_task_ignored() or with ANY other code which checks this task's flags.

I think this is WONTFIX.

Oleg.

#3 lore raw
Quoted 35 lines · Oleg Nesterov
Quoted 7 lines · Christian Brauner

This traces back to Oleg fixing a race between a group stop and a thread exiting before it notices that it has a pending signal or is in the middle of do_exit() already, causing group stop to get wacky.
The original commit to fix this race is
commit d12619b5ff56 ("fix group stop with exit race") which took sighand lock before setting PF_EXITING on the thread.

Not really... sig_task_ignored() didn't check task->flags until the recent 33da8e7c81 ("signal: Allow cifs and drbd to receive their terminating signals"). But I think this doesn't matter, see below.

Quoted 8 lines

If the race really matters and given how tsk->flags is currently accessed everywhere the simple fix for now might be:

kernel/signal.c +0 -0
index c4da1ef56fdf..cf61e044c4cc 100644

kernel/signal.c void exit_signals(struct task_struct *tsk) · line 2824

Quoted 6 lines
kernel/signal.c +2 -0
@@ -2819,7 +2819,9 @@ void exit_signals(struct task_struct *tsk)
        cgroup_threadgroup_change_begin(tsk);
 
        if (thread_group_empty(tsk) || signal_group_exit(tsk->signal)) {
+               spin_lock_irq(&tsk->sighand->siglock);
                tsk->flags |= PF_EXITING;
+               spin_unlock_irq(&tsk->sighand->siglock);

Well, exit_signals() tries to avoid ->siglock in this case....

But this doesn't matter. IIUC the problem is not that exit_signals() sets PF_EXITING, the problem is that it writes to tsk->flags and kasan detects the data race.

Quoted 5 lines

For example, freezable_schedule() which sets PF_FREEZER_SKIP can equally "race" with sig_task_ignored() or with ANY other code which checks this task's flags.

I think this is WONTFIX.

If taking the spinlock is unnecessary (which AFAIK it probably is) and there are no other writers to this flag, you will still need a WRITE_ONCE(tsk->flags, tsk->flags | PF_EXITING) to avoid the data-race.

However, if it is possible that there are concurrent writers setting other bits in flags, you need to ensure that the bits are set atomically (unless it's ok to lose some bits here). This can only be done via 1) taking siglock, or 2) via e.g. atomic_or(...), however, flags is not atomic_t ...

-- Marco

#4 lore raw
Quoted 9 lines · Marco Elver
Quoted 2 lines · Oleg Nesterov

I think this is WONTFIX.

If taking the spinlock is unnecessary (which AFAIK it probably is) and there are no other writers to this flag, you will still need a WRITE_ONCE(tsk->flags, tsk->flags | PF_EXITING) to avoid the data-race.

Or even WRITE_ONCE(tsk->flags, READ_ONCE(tsk->flags) | PF_EXITING) in theory. But in practice, I do not think compiler can turn

	curent->flags |= PF_EXITING;

into something which temporary clears another flag, say, PF_KTHREAD.

Quoted 2 lines

However, if it is possible that there are concurrent writers setting other bits in flags,

No, only current taks should change its ->flags.

Oleg.

#5 lore raw
In reply to Oleg Nesterov
Quoted 30 lines · Oleg Nesterov
Quoted 7 lines · Christian Brauner

This traces back to Oleg fixing a race between a group stop and a thread exiting before it notices that it has a pending signal or is in the middle of do_exit() already, causing group stop to get wacky.
The original commit to fix this race is
commit d12619b5ff56 ("fix group stop with exit race") which took sighand lock before setting PF_EXITING on the thread.

Not really... sig_task_ignored() didn't check task->flags until the recent 33da8e7c81 ("signal: Allow cifs and drbd to receive their terminating signals"). But I think this doesn't matter, see below.

Quoted 8 lines

If the race really matters and given how tsk->flags is currently accessed everywhere the simple fix for now might be:

kernel/signal.c +0 -0
index c4da1ef56fdf..cf61e044c4cc 100644

kernel/signal.c void exit_signals(struct task_struct *tsk) · line 2824

Quoted 6 lines
kernel/signal.c +2 -0
@@ -2819,7 +2819,9 @@ void exit_signals(struct task_struct *tsk)
        cgroup_threadgroup_change_begin(tsk);
 
        if (thread_group_empty(tsk) || signal_group_exit(tsk->signal)) {
+               spin_lock_irq(&tsk->sighand->siglock);
                tsk->flags |= PF_EXITING;
+               spin_unlock_irq(&tsk->sighand->siglock);

Well, exit_signals() tries to avoid ->siglock in this case....

Quoted 3 lines

But this doesn't matter. IIUC the problem is not that exit_signals() sets PF_EXITING, the problem is that it writes to tsk->flags and kasan detects the data race.

Right, that's the reason I said "If the race really matters". I thought that other writers/readers always take sighand lock. So the easy fix would have been to take sighand lock too.
The alternative is that we need to fiddle with task_struct itself and replace flags with an atomic_t or sm which is more invasive and probably more controversial.

Christian

#6 lore raw

cgroup_enable_task_cg_lists && PF_EXITING (Was: KCSAN: data-race in exit_signals / prepare_signal)

In reply to syzbot

could you explain the usage of siglock/PF_EXITING in cgroup_enable_task_cg_lists() ?

PF_EXITING is protected by cgroup_threadgroup_rwsem, not by sighand->siglock.

Oleg.

maintainer
#7 lore raw

Re: cgroup_enable_task_cg_lists && PF_EXITING (Was: KCSAN: data-race in exit_signals / prepare_signal)

Hello, Oleg.

Quoted 5 lines · Oleg Nesterov

could you explain the usage of siglock/PF_EXITING in cgroup_enable_task_cg_lists() ?

PF_EXITING is protected by cgroup_threadgroup_rwsem, not by sighand->siglock.

Yeah, the optimization was added a really long time ago and I'm not sure it was ever correct. I'm removing it. If this ever becomes a problem (pretty unlikely), I think the right thing to do is adding a boot param instead of trying to do this dynamically.

Thanks.

Signature · 1 line
tejun
maintainer
#8 lore raw

PATCH 1/1 cgroup: remove cgroup_enable_task_cg_lists() optimization

In reply to Oleg Nesterov

cgroup_enable_task_cg_lists() is used to lazyily initialize task cgroup associations on the first use to reduce fork / exit overheads on systems which don't use cgroup. Unfortunately, locking around it has never been actually correct and its value is dubious given how the vast majority of systems use cgroup right away from boot.

This patch removes the optimization. For now, replace the cg_list based branches with WARN_ON_ONCE()'s to be on the safe side. We can simplify the logic further in the future.

---
 include/linux/cgroup.h |    1 
 kernel/cgroup/cgroup.c |  184 ++++++++++---------------------------------------
 kernel/cgroup/cpuset.c |    2 
 3 files changed, 39 insertions(+), 148 deletions(-)
include/linux/cgroup.h +0 -1
index 3ba3e6da13a6..f6b048902d6c 100644
@@ -150,7 +150,6 @@ struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
 struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
 					struct cgroup_subsys_state **dst_cssp);
 
-void cgroup_enable_task_cg_lists(void);
 void css_task_iter_start(struct cgroup_subsys_state *css, unsigned int flags,
 			 struct css_task_iter *it);
 struct task_struct *css_task_iter_next(struct css_task_iter *it);
kernel/cgroup/cgroup.c +39 -145

238 lines not on this page · raw message

kernel/cgroup/cpuset.c +0 -2
index c52bc91f882b..faff8f99e8f2 100644
@@ -928,8 +928,6 @@ static void rebuild_root_domains(void)
 	lockdep_assert_cpus_held();
 	lockdep_assert_held(&sched_domains_mutex);
 
-	cgroup_enable_task_cg_lists();
-
 	rcu_read_lock();
 
 	/*
maintainer
#9 lore raw

Re: [PATCH cgroup/for-5.5] cgroup: remove cgroup_enable_task_cg_lists() optimization

Quoted 6 lines · Tejun Heo

cgroup_enable_task_cg_lists() is used to lazyily initialize task cgroup associations on the first use to reduce fork / exit overheads on systems which don't use cgroup. Unfortunately, locking around it has never been actually correct and its value is dubious given how the vast majority of systems use cgroup right away from boot.

Quoted 6 lines

This patch removes the optimization. For now, replace the cg_list based branches with WARN_ON_ONCE()'s to be on the safe side. We can simplify the logic further in the future.

Applying to cgroup/for-5.5.

Thanks.

Signature · 1 line
tejun
#10 lore raw

Re: [PATCH cgroup/for-5.5] cgroup: remove cgroup_enable_task_cg_lists() optimization

[+Dmitry]

Quoted 7 lines · Tejun Heo
Quoted 6 lines · Tejun Heo

cgroup_enable_task_cg_lists() is used to lazyily initialize task cgroup associations on the first use to reduce fork / exit overheads on systems which don't use cgroup. Unfortunately, locking around it has never been actually correct and its value is dubious given how the vast majority of systems use cgroup right away from boot.

Quoted 8 lines
Quoted 6 lines

This patch removes the optimization. For now, replace the cg_list based branches with WARN_ON_ONCE()'s to be on the safe side. We can simplify the logic further in the future.

Applying to cgroup/for-5.5.

The code you removed was the only place where task->flags was set from !current. So I think this fixes the syzbot data-race report in:
https://lore.kernel.org/r/[email protected]

Thanks!
Christian

#11 lore raw

Re: [PATCH cgroup/for-5.5] cgroup: remove cgroup_enable_task_cg_lists() optimization

Quoted 15 lines · Christian Brauner

[+Dmitry]

Quoted 11 lines · Tejun Heo
Quoted 10 lines · Tejun Heo

cgroup_enable_task_cg_lists() is used to lazyily initialize task cgroup associations on the first use to reduce fork / exit overheads on systems which don't use cgroup. Unfortunately, locking around it has never been actually correct and its value is dubious given how the vast majority of systems use cgroup right away from boot.

This patch removes the optimization. For now, replace the cg_list based branches with WARN_ON_ONCE()'s to be on the safe side. We can simplify the logic further in the future.

Quoted 7 lines
Quoted 4 lines
Quoted 2 lines

Applying to cgroup/for-5.5.

The code you removed was the only place where task->flags was set from !current.

No, that code doesn't modify task->flags. It checks PF_EXITING under siglock but this makes no sense and can't avoid the race with cgroup_exit().

Quoted 2 lines

So I think this fixes the syzbot data-race report in:
https://lore.kernel.org/r/[email protected]

No.

Almost every usage of task->flags (load or sore) can be reported as "data race".

Say, you do

	if (task->flags & PF_KTHREAD)

while this task does

	current->flags |= PF_FREEZER_SKIP;
	schedule().

this is data race.

Oleg.

#12 lore raw

Re: [PATCH cgroup/for-5.5] cgroup: remove cgroup_enable_task_cg_lists() optimization

Quoted 19 lines · Oleg Nesterov
Quoted 18 lines · Christian Brauner

[+Dmitry]

Quoted 14 lines · Tejun Heo
Quoted 12 lines · Tejun Heo

cgroup_enable_task_cg_lists() is used to lazyily initialize task cgroup associations on the first use to reduce fork / exit overheads on systems which don't use cgroup. Unfortunately, locking around it has never been actually correct and its value is dubious given how the vast majority of systems use cgroup right away from boot.

This patch removes the optimization. For now, replace the cg_list based branches with WARN_ON_ONCE()'s to be on the safe side. We can simplify the logic further in the future.

Quoted 7 lines
Quoted 4 lines
Quoted 1 line

Applying to cgroup/for-5.5.

The code you removed was the only place where task->flags was set from !current.

No, that code doesn't modify task->flags. It checks PF_EXITING under siglock but this makes no sense and can't avoid the race with cgroup_exit().

Sorry, you are right. I misread
Ah right, sorry I misremembered this from the prior thread where we discussed where ->flags is set from [1].

Quoted 10 lines
Quoted 3 lines

So I think this fixes the syzbot data-race report in:
https://lore.kernel.org/r/[email protected]

No.

Almost every usage of task->flags (load or sore) can be reported as "data race".

Say, you do

Quoted 8 lines
	if (task->flags & PF_KTHREAD)

while this task does

	current->flags |= PF_FREEZER_SKIP;
	schedule().

this is data race.

Right, but I thought we agreed on WONTFIX in those scenarios? The alternative is to READ_ONCE()/WRITE_ONCE() all of these.

[1]: https://lore.kernel.org/r/[email protected]

Anyway, accidental noise on my part.
Christian

#13 lore raw

Re: [PATCH cgroup/for-5.5] cgroup: remove cgroup_enable_task_cg_lists() optimization

Quoted 10 lines · Christian Brauner
Quoted 8 lines · Oleg Nesterov

Almost every usage of task->flags (load or sore) can be reported as "data race".

Say, you do

	if (task->flags & PF_KTHREAD)

while this task does

Quoted 7 lines
Quoted 4 lines
	current->flags |= PF_FREEZER_SKIP;
	schedule().

this is data race.

Right, but I thought we agreed on WONTFIX in those scenarios? The alternative is to READ_ONCE()/WRITE_ONCE() all of these.

Well, in my opinion this is WONTFIX, but I won't argue if someone adds _ONCE to all of these. Same for task->state, exit_state, and more.

Oleg.

#14 lore raw

Re: [PATCH cgroup/for-5.5] cgroup: remove cgroup_enable_task_cg_lists() optimization

[Removing a few people from Cc to avoid spamming the whole world]

Quoted 14 lines · Oleg Nesterov
Quoted 13 lines · Christian Brauner
Quoted 11 lines · Oleg Nesterov

Almost every usage of task->flags (load or sore) can be reported as "data race".

Say, you do

	if (task->flags & PF_KTHREAD)

while this task does

	current->flags |= PF_FREEZER_SKIP;
	schedule().
Quoted 8 lines
Quoted 4 lines
Quoted 1 line

this is data race.

Right, but I thought we agreed on WONTFIX in those scenarios? The alternative is to READ_ONCE()/WRITE_ONCE() all of these.

Well, in my opinion this is WONTFIX, but I won't argue if someone adds _ONCE to all of these. Same for task->state, exit_state, and more.

Well, I honestly think that state and exit_state would make sense. There already were issues that got fixed for example in 3245d6acab98 ("exit: fix race between wait_consider_task() and wait_task_zombie()") and as far as I understand this would also help kcsan to better detect races.

Christian

#15 lore raw

Re: [PATCH cgroup/for-5.5] cgroup: remove cgroup_enable_task_cg_lists() optimization

Quoted 18 lines · Christian Brauner
Quoted 16 lines · Oleg Nesterov
Quoted 15 lines · Christian Brauner
Quoted 12 lines · Oleg Nesterov

Almost every usage of task->flags (load or sore) can be reported as "data race".

Say, you do

	if (task->flags & PF_KTHREAD)

while this task does

	current->flags |= PF_FREEZER_SKIP;
	schedule().

this is data race.

Quoted 8 lines
Quoted 6 lines
Quoted 2 lines

Right, but I thought we agreed on WONTFIX in those scenarios? The alternative is to READ_ONCE()/WRITE_ONCE() all of these.

Well, in my opinion this is WONTFIX, but I won't argue if someone adds _ONCE to all of these. Same for task->state, exit_state, and more.

Well, I honestly think that state and exit_state would make sense.

Heh. Again, I am not arguing, but...

OK, lets suppose we blindly add READ_ONCE() to every access of task->state/exit_state.

Yes, this won't hurt and possibly can fix some bugs we are not aware of.

However,

Quoted 2 lines

There already were issues that got fixed for example in 3245d6acab98 ("exit: fix race between wait_consider_task() and wait_task_zombie()")

The change above can't fix the problem like this.

It is not that this code lacked READ_ONCE(). I am sure me and others understood that this code can read ->exit_state more than once, just nobody noticed that in this case this is really wrong.

IOW, if we simply change the code before 3245d6acab98 to use READ_ONCE() the code will be equally wrong, and

Quoted 2 lines

and as far as I understand this would also help kcsan to better detect races.

this change will simply hide the problem from kcsan.

Oleg.

#16 lore raw

Re: [PATCH cgroup/for-5.5] cgroup: remove cgroup_enable_task_cg_lists() optimization

Quoted 26 lines · Oleg Nesterov
Quoted 25 lines · Christian Brauner
Quoted 22 lines · Oleg Nesterov
Quoted 10 lines · Christian Brauner
Quoted 8 lines · Oleg Nesterov

Almost every usage of task->flags (load or sore) can be reported as "data race".

Say, you do

	if (task->flags & PF_KTHREAD)

while this task does

Quoted 7 lines
Quoted 4 lines
	current->flags |= PF_FREEZER_SKIP;
	schedule().

this is data race.

Right, but I thought we agreed on WONTFIX in those scenarios? The alternative is to READ_ONCE()/WRITE_ONCE() all of these.

Well, in my opinion this is WONTFIX, but I won't argue if someone adds _ONCE to all of these. Same for task->state, exit_state, and more.

Quoted 8 lines
Quoted 1 line

Well, I honestly think that state and exit_state would make sense.

Heh. Again, I am not arguing, but...

OK, lets suppose we blindly add READ_ONCE() to every access of task->state/exit_state.

Yes, this won't hurt and possibly can fix some bugs we are not aware of.

I wasn't planning or working on adding *_ONCE everywhere. ;) I just think it makes sense as a preemptive strike since they are shared (though mostly protected by locks anyway).

Quoted 7 lines

However,

Quoted 2 lines

There already were issues that got fixed for example in 3245d6acab98 ("exit: fix race between wait_consider_task() and wait_task_zombie()")

The change above can't fix the problem like this.

No argument about the code we discussed right here, for sure!

Quoted 5 lines

It is not that this code lacked READ_ONCE(). I am sure me and others understood that this code can read ->exit_state more than once, just nobody noticed that in this case this is really wrong.

Quoted 7 lines

IOW, if we simply change the code before 3245d6acab98 to use READ_ONCE() the code will be equally wrong, and

Quoted 2 lines

and as far as I understand this would also help kcsan to better detect races.

this change will simply hide the problem from kcsan.

I can't speak to that since the claim that read_once() helps them even if it's not really doing anything. But maybe I misunderstood the k{c,t}san manpage.

Christian

#17 lore raw

Re: [PATCH cgroup/for-5.5] cgroup: remove cgroup_enable_task_cg_lists() optimization

Quoted 10 lines · Christian Brauner
Quoted 9 lines · Oleg Nesterov

On 10/25, Christian Brauner wrote:

It is not that this code lacked READ_ONCE(). I am sure me and others understood that this code can read ->exit_state more than once, just nobody noticed that in this case this is really wrong.

IOW, if we simply change the code before 3245d6acab98 to use READ_ONCE() the code will be equally wrong, and

Quoted 8 lines
Quoted 4 lines
Quoted 2 lines

and as far as I understand this would also help kcsan to better detect races.

this change will simply hide the problem from kcsan.

I can't speak to that since the claim that read_once() helps them even if it's not really doing anything. But maybe I misunderstood the k{c,t}san manpage.

What Oleg said is right: marking things _ONCE will make these accesses no longer be data races, and KCSAN would then no longer report these as data race, even if the code is still buggy due to there being a race condition. In this case, the race condition would stop also being a data race, and we'd no longer find it.

Thanks,
-- Marco